4 ways CISOs can manage AI use in the enterprise

Over the summer, I wrote a column about how CIOs are worried about the informal rise of generative AI in the enterprise. That column may have been the understatement of the year.

Since then, many CIOs I’ve spoken with have grappled with enterprise data security and privacy issues around AI usage in their companies. A primary fear is that employees, partners, and organizational stakeholders might share everything from private data to source code into public large language models (LLMs), expose proprietary information and intellectual property, or reveal vulnerabilities to exploit. Other fears cover compliance with emerging AI regulations and the risk of models becoming contaminated or biased through adversarial attacks.

One Fortune 500 executive recently told me that they were worried that one organization could use a public LLM to learn what its competitors were asking that same LLM. For example, one pharmaceutical company using ChatGPT4 or similar for corporate espionage could essentially spy on its competitor’s research queries. A public LLM aggregates the data of the prompt itself and which user initiated that prompt. So, by asking about a certain company’s research, that data can become part of the public record.

But with time, CIOs are starting to figure out ways to manage the use of generative AI within the enterprise. A recent CIO column suggested that the biggest worry for CIOs should not be the fear of AI growth but rather figuring out the best way to gradually incorporate generative AI into the enterprise, either as an add-on model or a foundational piece of the architecture. While it is a given that AI will help organizations drive competitive advantage, the roll-out of this quickly evolving technology must be done safely.

Let’s dig a little deeper into how CIOs can succeed at helping manage the safe use of generative AI within their organizations.

Protecting data

In a recent meeting I attended with over 100 security executives, the prevailing theme among participants was that the primary techniques used today to manage the safe use of AI in their organization were employee training and usage policies. Most of these executives are primarily concerned with all of the bad things that can happen, as one might expect, not the advantages that AI can provide. A few even suggested that they have considered policies to prevent the use of public LLMs.

But conversely, trying to prohibit the use of and blocking of AI at the firewall would be akin to being considered a Luddite from the Stone Age – you simply cannot prevent access to AI and be a player in the 21^st century.

So, simply blocking LLM access is not the right answer. As some executives suggested, it would benefit the enterprise more to monitor the AI traffic at an aggregate (not employee-specific) level to understand the risks of public LLM usage. With that knowledge and training, policies can be better optimized to protect the organization while providing the advantages provided by AI.

Government oversight

All the recent innovation has certainly caught the U.S. government’s attention. And that’s why in October, the White House issued its guidelines on regulating AI in government agencies. The Executive Order calls for AI governance to move forward with urgency, with calls to start implementation in 90 days to a full year. While new laws around AI are likely still far off, I think it’s the right move for federal agencies to start shaping AI regulation for the broader market.

There is a solid argument that regulation will stifle innovation, particularly in this early phase of AI development. But, as any observer can see, it takes the government a long time to pass any laws relating to industry regulation so it’s time to get people talking and thinking about it now. AI is evolving faster than any tech wave we have seen in the past.

4 focus areas for CISOs

As a corporate CISO, it’s your responsibility to help manage the safe use of generative AI within your organization to protect your company. Here are four steps to take now as the industry, technologies, and regulations evolve.

1) Training, policy & process

Today, the most practical thing everyone can do is undergo AI training and implement company policies and processes. Just like we train people around phishing, ransomware, and other security topics, we need to train employees, partners, and other stakeholders about how AI works, the risks within the enterprise, how to use it sensibly, and how it may benefit (or potentially harm) the enterprise.

AI can create as much havoc (or more!) as any traditional phishing attack over the last five to 10 years. So effective training, empowered with AI policy and processes, is necessary for AI to gradually move into the enterprise.

2) Sandbox the public LLM

The second step for enterprise CISOs to consider is to sandbox the public LLMs. One CISO I spoke with recently has created a sandbox for seven LLMs to allow those connected to the network to harness the wisdom of the public LLM but without the reciprocal sharing of local knowledge back to the public LLM. The prompts can be answered in that sandbox, but the data requested in the prompt itself never gets sent back to OpenAI, or other AI innovators.

Another method to achieve this is to download the open-source LLMs and use them locally. Using an internal sandbox to reign in proprietary information may be worth pursuing.

3) Monitoring AI traffic

CISOs might start monitoring their network traffic to see which prompts are going out to LLMs and which information is coming back. Right now, it does not involve policing the networks and the information, but rather, it’s a watching, learning, and monitoring process, helping the security team make better decisions on how to safely use LLMs.

4) Future of LLM firewall (proxy server)

The LLMs could provide some sort of firewall capability similar to what we have now with the major cloud platforms. LLMs could take this functionality and make it a feature of their product for added security, incorporating an organization’s processes and policies.

If this functionality for firewalling or proxying emerges, there will be room for third-party providers, just as we’ve seen independent cloud security companies emerge to manage AWS, Azure and GCP. I’m not sure we’re at that stage yet, but with some sort of proxy server around LLMs, CIOs can more safely use generative AI within their organizations.

C-suite executives must know how AI works to truly gain enterprise value in using AI. AI’s speedy evolution and daily use cases are arriving at an accelerated pace. Leaders must acknowledge the risks and uncertainty of AI use in their organization as it is used to drive potential business value.

As I wrote previously, it’s ultimately necessary for organizations to chart the path for AI in the workplace. Whether the use of AI happens with stronger network security measures or by limiting employee use with internal controls, it’s mandatory that CISOs and their teams work with senior management on the four areas above to manage generative AI in the enterprise.

One final note. We are starting to see a new title joining enterprise executive teams – the Chief AI Officer (CAIO). This role, which would work alongside the CISO, could blossom into an area holding responsibilities for overall practical and marketable use of AI in the enterprise, helping CISOs ensure data security and AI process use.